Robustness in Point-based 3D Object Detection

Document Type



Thanks to developments in 3D data collecting and deep learning techniques, substantial progress has recently been made in 3D object detection on point clouds. However, 3D scenes are susceptible to sensor errors as well as pre-processing information loss. Designing techniques that are resistant to these variations is therefore crucial. This requires a detailed analysis and understanding of the effect of such variations. This thesis seeks to design, evaluate and compare a number of data corruptions to popular point-based 3D object detectors. To the best of our knowledge, we are the first to investigate the robustness of point-based 3D object detectors. To this end, we design and evaluate corruptions that involve data alteration. Furthermore, we use our corruptions and some other operations as data augmentations to boost the robustness of the models. More specifically, we automate the usage of the data augmentations during training and propose a diversity measure to control the choice of augmentations to impose a regularization effect which lead to better generalization and performance. We validate in experiments the effectiveness of our proposed measure and report performance gain of 1-2% mAP on the baseline of trained detectors models as well as we boost the model’s robustness.

First Page


Last Page


Publication Date



Thesis submitted to the Deanship of Graduate and Postdoctoral Studies

In partial fulfillment of the requirements for the M.Sc degree in Computer Vision

Advisors: Dr. Rao Anwer, Dr. Hisham Cholakkal

Online access for MBZUAI patrons